12

I have a list outputs from a sigmoid function as a tensor in PyTorch

E.g

output (type) = torch.Size([4]) tensor([0.4481, 0.4014, 0.5820, 0.2877], device='cuda:0',

As I'm doing binary classification I want to turn all values bellow 0.5 to 0 and above 0.5 to 1.

Traditionally with a NumPy array you can use list iterators:

output_prediction = [1 if x > 0.5 else 0 for x in outputs ]

This would work, however I have to later convert output_prediction back to a tensor to use

torch.sum(ouput_prediction == labels.data)

Where labels.data is a binary tensor of labels.

Is there a way to use list iterators with tensors?

Brian Formento
  • 731
  • 2
  • 9
  • 24
  • Maybe you could just use `output_prediction = torch.tensor([1 if x > 0.5 else 0 for x in outputs ])`? – lahsuk Sep 19 '19 at 04:27

3 Answers3

34
prob = torch.tensor([0.3,0.4,0.6,0.7])

out = (prob>0.5).float()
# tensor([0.,0.,1.,1.])

Explanation: In pytorch, you can directly use prob>0.5 to get a torch.bool type tensor. Then you can convert to float type via .float().

zihaozhihao
  • 4,197
  • 2
  • 15
  • 25
0

Why not consider using a loopless solution? Maybe something like below would suffice:

In [34]: output = torch.tensor([0.4481, 0.4014, 0.5820, 0.2877]) 

# subtract off the threshold value (0.5), create a boolean mask, 
# and then cast the resultant tensor to an `int` type
In [35]: result = torch.as_tensor((output - 0.5) > 0, dtype=torch.int32) 

In [36]: result        
Out[36]: tensor([0, 0, 1, 0], dtype=torch.int32)
kmario23
  • 57,311
  • 13
  • 161
  • 150
  • Do you have any insights on how the subtraction could add to the execution cost? Would it take longer than the accepted solution in your opinion? – dennlinger Sep 19 '19 at 07:40
  • 1
    @dennlinger Based on my tests, the time difference seems negligible or sometimes I even get no difference in execution time. – kmario23 Sep 19 '19 at 08:00
0
result = torch.as_tensor((output - 0.5) > 0, dtype=torch.int32), turns the require_grad to False.
To train your model use this code:
<p>>m = torch.nn.Sigmoid()</p>
>loss = criterion(m(output),target)

review above code.

Rajat Jaiswal
  • 645
  • 4
  • 15
Soham Mitra
  • 1